Computation Cost-Driven Offloading Strategy Based on Reinforcement Learning for Consumer Devices

Rui Zhang, Hui Xia, Zijun Chen, Zi Kang, Kai Wang, Wei Gao

Abstract: Edge computing extends computational capabilities to the physical edge of the world, providing faster and more efficient responses to consumer devices such as smart home appliances, smartphones, and wearable devices. However, it raises a critical question: How can we efficiently manage time-sensitive and computationally intensive tasks to meet device performance needs? To address this, we propose a computational cost-driven strategy for computation offloading based on reinforcement learning for consumer devices. Firstly, we convert the problem of minimizing computational costs, including latency and energy consumption, into maximizing the cumulative rewards problem of consumer devices. Second, we design a task completion time estimation method to ensure a favorable user experience. Third, by comprehensively considering the status of the network, computational demands, and task completion times, we employ reinforcement learning techniques to determine the optimal task offload strategy for consumer devices by maximizing cumulative rewards. This ensures efficient task completion with lower computational costs. The simulation results demonstrate that our method achieves the highest rewards in different scenarios. Our method reduces the task discard ratio by 25% compared to the PPO_KLP-based offload strategy.